Accessible, uniform protein property prediction with a scikit-learn based toolset AIDE

Evan Komp, Kristoffer E. Johansson, Nicholas P. Gauthier, Japheth E. Gado, Kresten Lindorff‐Larsen, Gregg T. BeckhamBioinformatics 2025 · DOI

TL;DR

Protein property prediction (PPP) is a rapidly growing field, but the tools and methods are scattered and often have disparate interfaces, making it hard to compare and use them. This paper introduces AIDE, a Python package that provides a unified, scikit-learn compatible API for a wide range of existing PPP tools, simplifying their integration into reproducible research workflows.

What Problem Does This Paper Address?

Proteins are the workhorses of biology, performing diverse functions essential for life. Understanding and predicting their properties (like stability, activity, or binding affinity) is crucial for fields ranging from drug discovery to enzyme engineering. However, the landscape of protein property prediction methods is fragmented. Researchers often have to navigate a multitude of tools, each with its own installation process, data input requirements, and API. This makes it challenging to:

  • Compare different methods: It's difficult to fairly assess which method performs best for a given task.
  • Reproduce results: Replicating previous work requires meticulously setting up the same environment and code for each tool.
  • Build complex pipelines: Combining multiple prediction steps (e.g., embedding a protein sequence and then predicting its property) becomes a significant engineering effort.
  • Onboard new researchers: The steep learning curve for each individual tool hinders broader adoption.

This paper aims to solve this by creating a standardized, accessible, and reproducible framework for protein property prediction.

Methods

The core of AIDE is its adherence to the scikit-learn API. This means that protein models within AIDE are designed to work like standard scikit-learn estimators, offering fit, transform, and predict methods. This allows users to seamlessly integrate AIDE models into existing scikit-learn pipelines, enabling complex workflows with minimal custom code.

Key methodological contributions include:

  • Modular Base Classes: AIDE provides base classes for protein data structures (sequences, alignments, structures) and protein models.
  • Mixin Protocols: These are designed to handle common requirements and behaviors of different protein models, such as needing a multiple sequence alignment (MSA) for fitting or requiring a wild-type sequence. This modularity simplifies the process of wrapping new prediction methods.
  • Drop-in Compatibility: By inheriting from scikit-learn classes, AIDE models can be used alongside standard scikit-learn transformers (like scalers or PCA) and estimators.
  • Support for Diverse Models: AIDE wraps both supervised learning (SML) methods and "zero-shot" (ZS) learning methods, which can predict properties without direct training data for that specific task.
  • Dependency Management: The package attempts to manage dependencies for different models, allowing users to install only what they need, preventing overly bloated environments. For computationally intensive models that require specific environments (like PyTorch or CUDA), AIDE can call them as subprocesses, maintaining compatibility.

Datasets

The paper doesn't introduce new benchmark datasets for training. Instead, it leverages existing datasets and benchmarks to showcase AIDE's capabilities.

  • Johnston et al. (2024) epistatic combinatorial enzyme activity dataset: Used in Showcase 1 to benchmark various ZS and supervised predictors.
  • PETase homolog activity dataset (Norton-Baker et al. 2025): Used in Showcase 2 to build an activity predictor for identifying novel enzymes.
  • ProteinGym: Mentioned as a benchmark for evaluating wrapped models, specifically the ENV_ECOLI benchmark.

Key Insights and Ideas

  • The Power of Standardization: The most significant insight is how much friction can be removed by adopting a widely-used, standardized API like scikit-learn's. This dramatically lowers the barrier to entry for using complex protein prediction models.
  • Bridging the Gap Between ZS and SML: AIDE facilitates the combination of "zero-shot" (predictive without direct training) and supervised learning methods, allowing researchers to augment ZS predictions with supervised models for potentially improved performance.
  • Reproducibility as a Cornerstone: By enabling the creation of reproducible scripts and pipelines, AIDE directly addresses a critical need in scientific research.
  • Community Contribution: The modular design and clear API are intended to encourage the community to contribute wrappers for new and existing PPP methods, fostering a collaborative ecosystem.

Key Results

The paper demonstrates AIDE's utility through two showcases and several examples:

  • Showcase 1: Successfully benchmarked various embedding strategies (one-hot, ESM, EVCouplings) and predictor types (linear, nonlinear) on an epistatic dataset. The entire process, including hyperparameter optimization and cross-validation, was completed on a laptop in under 8 hours.
  • Showcase 2: Built a PETase activity predictor using different embedding strategies and predictor heads. This was achieved in a single notebook with a runtime of under 4 hours.
  • Unit Tests: AIDE itself is extensively unit tested, achieving over 80% code coverage as of September 2025.
  • Compatibility: The framework is designed to be compatible with Windows, Mac, and Linux.

Limitations and Future Work

  • GPU Dependency: Many of the wrapped models require a GPU for efficient operation, and some assume CUDA. This is a practical limitation for users without access to such hardware.
  • Testing Wrapped Models: While AIDE's core is tested, the paper acknowledges that it cannot guarantee the correctness of the wrapped external methods themselves.
  • Expanding Model Support: The current list of wrapped tools is presented as a "proof of concept." The authors explicitly plan to add more models and data modes, encouraging community contributions.
  • Mixins for New Models: As the PPP field evolves, more Mixins will likely be needed to accommodate novel model architectures and requirements.

Summary updated

Abstract

AIDE is an installable, importable python package inheriting from scikit-learn classes and API and is installable on Windows, Mac, and Linux. Many of the wrapped models internal to AIDE will be effectively inaccessible without a GPU, and some assume CUDA. The newest stable, tested version can be found at https://github.com/beckham-lab/aide_predict and a full user guide and API reference can be found at https://beckham-lab.github.io/aide_predict/. Static versions of both at the time of writing can be found on Zenodo.
BibTeX
@article{komp2025accessible,
  author = {Evan Komp and Kristoffer E. Johansson and Nicholas P. Gauthier and Japheth E. Gado and Kresten Lindorff‐Larsen and Gregg T. Beckham},
  title = {{Accessible, uniform protein property prediction with a scikit-learn based toolset AIDE}},
  journal = {Bioinformatics},
  year = {2025},
  doi = {10.1093/bioinformatics/btaf544},
}
Figure